Surgical planning for head and neck squamous cell carcinoma (HNSCC) requires accurate delineation of surgical target volumes (STVs), which encompass lymphatic and fibrofatty tissue at risk for metastasis. Current machine learning models focus largely on radiotherapy structures and do not capture surgically relevant tissues. This study developed a 2.5D U-Net model trained on 40 contrast-enhanced CT scans from DHMC and TCGA-HNSCC datasets using a hybrid uncertainty-guided active learning and semi-supervised learning pipeline to overcome limited labeled data. On an independent test set of 20 cases, the model generated STVs with a mean Dice score of 0.72 ± 0.06, showing strong concordance with human segmentation despite occasional under-segmentation and missed nodal disease. Manual STVs correlated with lymphadenectomy specimen mass and postoperative lymph node counts, supporting their biological relevance. While the trained model demonstrated promising early performance, it tended to under-predict volume, indicating the need for additional training data and refined uncertainty modeling. Future directions include multi-reader evaluations, model refinement using Monte Carlo dropout, and assessment of clinical utility for surgical workflow optimization.